Prompt Injection Security Policy

Defines controls to mitigate prompt injection risks, particularly in agentic systems like WSAI CODE.

All products processing untrusted external content within model context.

Prompt Injection.

⟳Prescriptive Technical Mandates, replacing generic language Systems ingesting untrusted content must implement, at minimum: - Privilege separation: Untrusted content is never granted the same instruction-authority as system or user prompts; a formal instruction hierarchy (system > developer > user > tool/retrieved-content) is enforced at the model and application layer. - Output filtering: All agent outputs pass through an action-classifier before execution; outputs matching known injection patterns are blocked and logged. - Scoped, revocable permissions: Agents receive time-boxed, task-scoped credentials rather than standing access; permissions expire automatically at task completion. - Mandatory penetration testing: Every new tool integration undergoes dedicated prompt injection penetration testing before release, performed by a team independent of the feature's builders. - Cryptographically signed audit logs: All agent actions are logged in a hash-chained, tamper-evident log reviewable during incident investigation.

Security and Research Engineering jointly own mitigation design; independent penetration testing team validates before release.

Prompt injection test suites run against all agentic feature releases; audit log integrity verified quarterly.

Features failing injection resistance testing blocked from release.

security@webspaceai.in | vuln@webspaceai.in

CISO + AI Safety Research Lead joint approval, time-limited to 30 days pending remediation.

Quarterly.

security@webspaceai.in